We treat P300 speller paradigm as a 2-class classification problem. The core task is to distinguish signals evoked by target stimuli from those by non-target stimuli. The prediction of character is based on scores of 12 codes, accumulated by 15 (or 5) repetitions. To classify signal after each intensification, we implement a pattern recognition method called BAGGING with component classifier LDA(Linear discriminant analysis). For each subject, we first create 150 training set by drawing about 60% samples form original training set. Then each of these data sets is used to train a LDA classifier. The final classification decision is based on the vote of each component classifier. Some important parameters are listed as follows: Filtering: all data are band-pass filtered between 0.5-15Hz; Channel Selection: The No. 34, 11, 51, 62, 9, 13, 49, 53, 56 and 60 channels are selected; Feature vector: Signals (lasting 900ms from stimulus) from above channels are concatenated, and then down-sampled to 1/8.